Improving Differentially Private Models with Active Learning